perf: remove Python overhead from model hot paths (#3374)
* perf: remove Python overhead from model hot paths Signed-off-by: Alexandros Koumparoulis <akoumparouli@nvidia.com> * fix: preserve hot-path semantics Signed-off-by: Alexandros Koumparoulis <akoumparouli@nvidia.com> * fix: preserve Rice attention device compatibility Signed-off-by: Alexandros Koumparoulis <akoumparouli@nvidia.com> * fix: bound hot-path signature caches Signed-off-by: Alexandros Koumparoulis <akoumparouli@nvidia.com> * fix(moe): preserve deterministic expert bias gradients Signed-off-by: Alexandros Koumparoulis <akoumparouli@nvidia.com> * refactor(moe): split bias optimization Signed-off-by: Alexandros Koumparoulis <akoumparouli@nvidia.com> --------- Signed-off-by: Alexandros Koumparoulis <akoumparouli@nvidia.com>
A
Alexandros Koumparoulis committed
e802f0a5094b93f946a125032aafac0832090430
Parent: 3155217
Committed by GitHub <noreply@github.com>
on 8/4/2026, 4:09:03 PM